> TL;DR: NanoGen is a unified framework that trains and evaluates diffusion transformers across ImageNet and text-to-image with only roughly 12 lines of config change. We use it to show that method ranking shows no strong correlation between ImageNet and T2I generation, and introduce DiffusionBench, a holistic benchmark for DiT research
With a "TL;DR":
> TL;DR: NanoGen is a unified framework that trains and evaluates diffusion transformers across ImageNet and text-to-image with only roughly 12 lines of config change. We use it to show that method ranking shows no strong correlation between ImageNet and T2I generation, and introduce DiffusionBench, a holistic benchmark for DiT research